xcp: An extended cp

xcp is a (partial) clone of the Unix cp command. It is not intended as a full replacement, but as a companion utility with some more user-friendly feedback and some optimisations that make sense under certain tasks (see below).

Warning: xcp is currently beta-level software and almost certainly contains bugs and unexpected or inconsistent behaviour. It probably shouldn't be used for anything critical yet.

Please note that there are some known issues with copying files from virtual filesystems (e.g. /proc, /sys). See this LWN article for an overview of some of the complexities of dealing with kernel-generated files. This is a common problem with file utilities which rely on random access; for example rsync has the same issue.

Installation

Cargo

xcp can be installed directly from crates.io with:

cargo install xcp

Arch Linux

xcp is available on the Arch Linux User Repository. If you use an AUR helper, you can execute a command such as this:

yay -S xcp

NetBSD

xcp is available on NetBSD from the official repositories. To install it, simply run:

pkgin install xcp

Features and Anti-Features

Features

• Displays a progress-bar, both for directory and single file copies. This can be disabled with --no-progress.
• On Linux it uses copy_file_range call to copy files. This is the most efficient method of file-copying under Linux; in particular it is filesystem-aware, and can massively speed-up copies on network mounts by performing the copy operations server-side. However, unlike copy_file_range sparse files are detected and handled appropriately.
• Support for modern filesystem features such as reflinks.
• Optimised for 'modern' systems (i.e. multiple cores, copious RAM, and solid-state disks, especially ones connected into the main system bus, e.g. NVMe).
• Optional aggressive parallelism for systems with parallel IO. Quick experiments on a modern laptop suggest there may be benefits to parallel copies on NVMe disks. This is obviously highly system-dependent.
• Switchable 'drivers' to facilitate experimenting with alternative strategies for copy optimisation. Currently 2 drivers are available:
  • 'parfile': the previous hard-coded xcp copy method, which parallelises tree-walking and per-file copying. This is the default.
  • 'parblock': An experimental driver that parallelises copying at the block level. This has the potential for performance improvements in some architectures, but increases complexity. Testing is welcome.
• Non-Linux Unix-like OSs (OS X, *BSD) are supported via fall-back operation (although sparse-files are not yet supported in this case).
• Optionally understands .gitignore files to limit the copied directories.
• Optional native file-globbing.

(Possible) future features

• Conversion of files to sparse where appropriate, as with cp's --sparse=always flag.
• Aggressive sparseness detection with lseek.
• On non-Linux OSs sparse-files are not currenty supported but could be added if supported by the OS.

Differences with cp

• Permissions, xattrs and ACLs are copied by default; this can be disabled with --no-perms.
• Virtual file copies are not supported; for example /proc and /sys files.
• Character files such as sockets and pipes are copied as devices (i.e. via mknod) rather than copying their contents as a stream.
• The --reflink=never option may silently perform a reflink operation regardless. This is due to the use of copy_file_range which has no such override and may perform its own optimisations.
• cp 'simple' backups are not supported, only numbered.
• Some cp options are not available but may be added in the future.

Performance

Benchmarks are mostly meaningless, but the following are results from a laptop with an NVMe disk and in single-user mode. The target copy directory is a git checkout of the Firefox codebase, having been recently gc'd (i.e. a single 4.1GB pack file). fstrim -va and echo 3 | sudo tee /proc/sys/vm/drop_caches are run before each test run to minimise SSD allocation performance interference.

Note: xcp is optimised for 'modern' systems with lots of RAM and solid-state disks. In particular it is likely to perform worse on spinning disks unless they are in highly parallel arrays.

Local copy

• Single 4.1GB file copy, with the kernel cache dropped each run:
  • cp: ~6.2s
  • xcp: ~4.2s
• Single 4.1GB file copy, warmed cache (3 runs each):
  • cp: ~1.85s
  • xcp: ~1.7s
• Directory copy, kernel cache dropped each run:
  • cp: ~48s
  • xcp: ~56s
• Directory copy, warmed cache (3 runs each):
  • cp: ~6.9s
  • xcp: ~7.4s

NFS copy

xcp uses copy_file_range, which is filesystem aware. On NFSv4 this will result in the copy occurring server-side rather than transferring across the network. For large files this can be a significant win:

• Single 4.1GB file on NFSv4 mount
  • cp: 6m18s
  • xcp: 0m37s